Claude Opus 5
An upcoming flagship model from Anthropic with capabilities approaching the previous generation's flagship level, but at significantly reduced inference cost. Its scheduled release is July 2025.
Core Facts
Anthropic称Opus 5是迄今最对齐的模型，遵守宪法程度超过Sonnet 5和Fable，欺骗行为率最低
Opus 5.5 reaches the level of Fable 5.1 on most tasks, but with approximately 40% lower operating costs and faster speed.
DeepSeek-V4.1-Flash-1017 scores 77.0 on DeepSWE v1.1, surpassing Claude Opus 5's 74.0.
Timeline (last 90 days)
In automated benchmark tests, MiniMax M3.1 Flash scored 62%, Claude Opus 5 scored 50.3%, and GPT-5.6 Sol scored 45.8%.
In a test to create a landing page for a hardware device launch, MiniMax spent only 1 cent to complete it, while Opus 5 required $1.22, making it nearly 88 times cheaper and also 4 times faster.
The optimizer models used in the GEPA-Curator experiments are Gemini 3.1 Pro and Claude Opus 5.
On the SWE Bench Verified benchmark, Claude Opus 5 ranks first with a score of 97%, and top models generally score between 93% and 97%.
On the Vending Bench 2 benchmark, GPT-6 Astra finished with $15,500, while Claude Opus 5 finished with $11,000.
According to third-party estimates, the input price of DeepSeek V4.1 Flash during weekday daytime is about one-thirteenth to one-seventeenth that of GPT-5.6 and Claude Opus 5, and the output price is one-seventeenth to one-twenty-first.
Using Claude Opus 5, Maestro autonomously discovered proofs for 34 of 48 theorems in the June 2026 version of arXivLean, whereas the native agent managed only 19.
Claude Opus 5.5 was launched on September 22, 2026, reaching the level of Claude Fable 5.1 on most work while costing 40% less to run than Opus 5.
In the correspondence rate with papers published after the model's cutoff date, DiscoveryRSI-27B reached 6.3%, GPT-6 was 2.5%, and Claude Opus 5 was 2.3%.
According to the creator's real-world tests and community feedback, DeepSeek-V4.1-flash has surpassed Claude Opus 5, GPT 5.6 Sol, and a range of domestic open-source models on core metrics measuring real-world programming capability.
40 more timeline events
All Facts (20)
Anthropic称Opus 5是迄今最对齐的模型，遵守宪法程度超过Sonnet 5和Fable，欺骗行为率最低
85%VerifiedOpus 5.5 reaches the level of Fable 5.1 on most tasks, but with approximately 40% lower operating costs and faster speed.
80%VerifiedDeepSeek-V4.1-Flash-1017 scores 77.0 on DeepSWE v1.1, surpassing Claude Opus 5's 74.0.
70%VerifiedAnthropic states that Opus 5.5 can reduce total costs for typical tasks by approximately 40% and improve output speed by more than 30%.
70%VerifiedV4.1-Flash scores 1 point higher than a GPT-5.6-level model in comparison.
65%VerifiedAccounting for real usage scenarios, the actual cost advantage of Opus 5 over Fable is approximately 20% to 25%, with Fable costing about $2.75 per task and Opus 5 about $2.03.
65%VerifiedClaude 3.5 Sonnet has approached or even surpassed the performance of the early Claude 3 Opus in most benchmark tests.
65%VerifiedClaude Opus 5 is Anthropic's flagship large language model released in 2025, representing the highest capability tier of the Claude series
60%VerifiedDevin has integrated Anthropic's Claude Opus 5 model, live in Devin Desktop and Devin CLI
65%Unverified能力提升型技能存在时效性问题，随着模型版本升级（如未来的Claude Opus 5），某些能力提升技能可能不再需要
70%UnverifiedIn the Automation Bench test, GPT-6 Sol's extra high tier achieved first place 33% of the time, while 5.6's max effort scored 28.8.
60%UnverifiedThe effort parameter controls how much the model thinks rather than how much it says; reducing effort does not reliably shorten visible responses, and controlling response length requires explicit prompting.
60%UnverifiedThe model is capable of synthesizing SVG vector visual elements on the fly rather than downloading assets from the internet.
60%UnverifiedClaude Opus 5 scored 30.2% on ARC-AGI-3, compared to only 1.5% for the previous generation.
60%UnverifiedClaude Opus 5 scored 1861 on the knowledge work benchmark GDPval, above the human average.
60%UnverifiedOpus 5 remains proficient at discovering vulnerabilities, but its ability to exploit them has been significantly weakened.
60%UnverifiedClaude Opus 5 approaches Fable-level performance on the FrontierCode 1.1 benchmark while costing only half as much
60%UnverifiedIn automated benchmark tests, MiniMax M3.1 Flash scored 62%, Claude Opus 5 scored 50.3%, and GPT-5.6 Sol scored 45.8%.
50%UnverifiedIn a test to create a landing page for a hardware device launch, MiniMax spent only 1 cent to complete it, while Opus 5 required $1.22, making it nearly 88 times cheaper and also 4 times faster.
50%UnverifiedThe optimizer models used in the GEPA-Curator experiments are Gemini 3.1 Pro and Claude Opus 5.
50%